Nvidia should quit phoney war with Intel

28 Apr 2008 | 14:52 BST

By Nebojsa Novakovic

EVER SINCE Jen Hsun Huang's tirade, so well reported all over the place, I've been wondering what was all the noise was about.

I don't prefer integrated graphics, and never have really liked the idea either, as the "intelligent I/O" approach for everything may be a more balanced choice for a PC desktop anyway - rather than stealing cycles from CPU and/or its memory buses to refresh frames and play with textures - or even sounds, RAID computations and so on.

Also, the current batch of Intel's integrated graphics chipsets, pre-G45, isn't stellar on either performance or features. Intel is working hard on the compatibility problems with specific games, but the real performance boost will only be seen in the G45, hopefully out in the next month or two.

So, instead of the battle, let's talk shop. Why didn't Nvidia simply make full use of Intel's G45 delay and launch its own, "better" integrated graphics chipset for all those Core 2 CPUs? Wouldn't that be a far sweeter way of trouncing the giant, with real money gains instead of an all-out public war with a far bigger opponent?

Right now, despite that belated Phenom B3 intro, Intel's desktop and notebook market share simply overwhelms AMD's. Also, AMD's current 780G series integrated graphics chipsets are stronger in the GPU department than the Intel G35, and therefore more competitive against any Nvidia chipset offerings for the AMD chips.

So, if Nvidia had done the Nforce IGP for Intel first this time, it would have had roughly quadruple the target market, and less competition within it too! But, due to whatever reasons, maybe just to show disliking towards Intel or who knows what, the focus went to AMD and its - still, even after B3 - rather limp CPUs.

By the time Intel does come out with the G45, its integrated GPU capabilities will jump quite a bit - whether it is really triple the 3-D performance and smooth real time Blu-ray HD, we'll see when it comes. Same for the improved DX10 compatibility with those tricky games. Whatever the case, it will be that much harder for NV to come in then.

Interestingly, Intel kept quiet after Nvidia's boss' tirade. The firm mostly kept quiet after similar, but not so intense, AMD outbursts before this. It simply continued delivering. The current CPU situation shows it all too well - even if Nehalem doesn't come out till October or November, there's nothing AMD can do. Unless its 45nm parts cross way above the 3GHz line at launch, of course. Yet, Intel pairs Nehalems with AMD graphics...

The real battle here is Larrabee. The recent GPU launch acceleration by both AMD and Nvidia is, in a way, a tussle to gain strong position ready for the new entry next year. AMD is the first to break 1 Tflop peak single precision FP performance with the HD4870 card - even more useful would be having a full IEEE754-compliant double-precision FP on that card, a rumoured new feature. Then you could also offload large spreadsheet computations to the card finally. And, yes a pair of HD4870X2 cards could hit four TFlops in CrossfireX.

Then, Nvidia Geforce 9900GTX is supposed to go above teraflop and a half per chip, or near 5 Tflops in those extremely rare Tri-SLI setups. And, it's now expected in July, not September. Of course, it's not just peak Tflops, but how much of them can be used, that matters.

Let's assume that both vendors have 40 nm - TSMC or otherwise - GPU shrinks of these two entries by mid-2009, when Larrabee should be about ready for launch. And, let's give each a generous 30% performance jump. So, we'd be talking of above 1.3 Tflops per AMD chip, and near 2 Tflops for the Nvidia.

Last year IDF "Kicking Pat" chip demo already showcased 2 Tflops on a die which, we assume, would be continuously improved as the Larrabee is finalised. I'd safely say that Intel can, if they wish, come out with 2 Tflop top-end Larrabee in mid-late 2009, and that the chip would support full IEEE FP on both SP and DP number formats.

Even if the initial graphics drivers aren't perfect, Intel's computational experience would most likely leave Nvidia in the dust - AMD has greater chances here since, well, they were doing computational stuff for a while, too.

What can AMD and NV do to gain more GPGPU space prior to Larrabee? If we were to see the full computation GPU capability enabled in all high-end 3-D cards by both companies, not just those customised, overpriced "Tesla" and "Firestream" kits, then they'd be able to get more app support prior to Larrabee launch. How about that?

Before "Laughabee" comes, as Jen Hsun would say - and anyone can see, that's the real thing on his mind, not so much the chipset rant - there is still a lot to be done. Like, new chipsets for those Intel platforms; NV can surely select and put in an even better graphics core from their old GeForce 8xxx portfolio, to exceed the G45 decisively - it is up to them. Yes it may affect some low-end add-on graphics cards, but well, you can't have it all.

Then, before Larrabee, there are also desktop Nehalems early next year. If Intel gives Nvidia a QPI license for chipsets, I believe NV would miss a lot if they don't get out an offering there - remember, most initial Nehalems DON'T have an integrated GPU, so there still is space for chipset GPU integration.

So, rather than fight a still distant enemy, why not bury the hatchet and get some stuff done together? Nvidia might make better sense to work with Intel on the ample integrated graphics chipset market for the Core 2, and the upcoming support for Nehalems. That might just lead to more high-end Geforce and Quadro cards being shown on the near-final Nehalem machines this August at the Fall IDF in old Frisco, together with the DAAMIT offerings of course. To me, that'll be a good sign of the "state of the affairs". µ

© 2007 Incisive Media Investments Ltd. 2007